面向多动症的人体行为表达模型和病情量化函数设计方法研究

Attention deficit hyperactivity disorder (ADHD) is a common behavioral disorder for children with a very high incidence (3%-7% worldwide). Since ADHD is a chronic disease with symptoms lasting for years or even lifetime, it is necessary to continuously observe ADHD patients’ behavior in daily life to monitor their condition and provide feedback to doctors. However, due to the subjective factors of the evaluators or the expensive diagnostic equipment, it is impossible to rely on family members or existing diagnostic methods. To solve this problem, this project aims to use patients’ daily-life video for ADHD behavior analysis to achieve continuous and objective condition monitoring. The research content mainly includes: based on the principle of video multi-event detection, this project will conduct the cotext event modeling of ADHD behavior; based on the body three-dimensional skeleton model, this project will design the the expression model applicable to ADHD behavior; under the situation that there is no prior knowledge about the relation between ADHD behavior features and the patients’ condition, the condition quantification function will be studied. The research results can provide theoretical and techniqual bases for the development of continuous and objective monitoring systems for ADHD patients, and will also be significant for the improvement of ADHD prevention and treatment.

注意缺陷多动障碍(俗称多动症)是儿童常见的以行为障碍为特征的综合症，发病率极高(全球3％～7％)。由于多动症多呈慢性过程，症状持续多年甚至终身，需要在日常生活中观察患儿多动症行为来持续监测其病情、反馈治疗效果。然而受限于评定者的主观因素或昂贵的诊断设备，无法依赖家人和现有的多动症诊断方法。针对此问题，本项目探索利用患儿视频进行多动症行为分析以实现持续、客观的病情监测，提出了多动症的行为表达模型和多动症病情量化函数的设计方法。研究内容主要包括：根据视频中多目标事件检测原理，研究多动症行为的背景事件建模方法；在背景事件下，基于人体三维骨骼模型研究多动症行为的表达模型；在未知多动症行为特征与病情的量化关系情况下，研究多动症病情量化函数的设计方法。本项目的研究成果可为多动症患者病情监测系统的研制提供理论和方法基础，对多动症防治水平的提高具有重要意义。

项目针对在日常生活中需持续、客观监测多动症患者病情的需求，研究了多动症行为的表达模型用于多动症识别和基于遗传规划的回归分析方法用于病情程度量化。项目已按预期目标顺利完成，主要研究成果包括：.(1)设计了基于视觉注意力的多动症行为表达方法来提取行为关键信息，结果证明该方法比现有行为识别方法具有更鲁棒、更优的多动症识别效果。.(2)设计了基于动作特征增强和边界匹配网络的多动症行为提名生成方法，使特征向量与行为起止边界建立更准确的联系，有助于对多动症行为实现准确检测。.(3)设计了多种混合遗传规划算法，将局部搜索引入遗传规划算法中，相比基准遗传规划算法能更快、更准确地完成回归任务。.(4)设计了多种多目标遗传规划算法，并提出多种多目标占优机制，有效解决了遗传规划算法的膨胀问题、多目标技术中最终解选择等问题，在回归任务中性能优于基准多目标遗传规划算法。.在项目的支持下，共发表SCI检索期刊论文 11 篇（其中SCI一区3篇，SCI二区6篇），申请发明专利3项。在项目的支持下，共培养硕士研究生5名。